On the Artificial Evolution of Neural Graph Grammars

摘要: Artificial neural networks and other connectionist models of computation are frequently credited with biological plausibility. Since systems products Darwinian evolution, network optimisation by artificial evolutions has considerable appeal. However, the computational expense this can become prohibitive unless an emphasis is placed on modularity reuse. Gene expression holds answer to this. Genes translated into proteins that self-organize phenotypic traits such as brain, feedback loops controlling further genes. In paper we present a generalization mechanism, context-free graph grammar describes finite-state automata. The generated replacing hyperedges subgraphs automata according set hypergraph productions. These need not be homogeneous, e.g. they may correspond different types neurons, reflecting diversity neurons in brain. Desirable productions retrieved from population productions, which evolve mutation existing subsequent selection against user-defined criterion.